Annie Louis & Ani Nenkova University of Pennsylvania

Annie Louis & Ani Nenkova

University of Pennsylvania

Evaluation using system output

2

Evaluation using only the input

• Predicting when systems would do badly

• Wisdom of the crowds

- all systems‘ output make a great model

• High input-summary similarity = better summary

• Adding pseudo-models to human models

Numbers we report are not the officially distributed

ones from the AESOP track

Two uberbaselines—human summaries were

included which invalidated the results computed

Correlations were recomputed

Only difference—uberbaselines excluded

3

44 multi-document inputs

2 tasks ◦ Query focused

◦ Update

53 automatic systems◦ 52 peers, 1 automatic baseline

2 oracle systems◦ Not used in our work

Pyramid evaluation◦ Multiple human summaries – 4 models in TAC ‘09

◦ Can provide feedback about why a summary is bad

◦ Significant annotation effort

5

Responsiveness scores◦ Combined measure of content and linguistic quality

◦ Direct human judgements

◦ Scale 1 - 10

6

System-level ~ which system is better overall?◦ Average predicted scores for a system over the test set

◦ Average human scores

◦ Correlation between rankings

Input-level ~ which summary is better for an input?◦ Correlation between rankings of summaries for each

individual input

◦ % of inputs with significant correlations

Evaluate content selection using no human

models at all

7

Evaluation on non-standard test sets◦ With no model summaries

Likely to be a good objective function for content

selection

But many ways to measure similarity◦ KL, JS divergence

◦ Cosine similarity

◦ Topic word similarity

◦ Frequency based summary likelihood

8

Analysis of different input-summary similarity

metrics [TAC ‗08, EMNLP ‗09]

9

Top features were highly predictive of human

scores◦ Best correlation at system-level ~ 0.89

Performance varies with different features◦ Best features ~ information-theoretic measures

◦ Worst ~ frequency based metrics

Distance between 2 probability distributions◦ As average KL divergence from their mean distribution

Low divergence ~ better summary quality

10

)]||()||([)||(2

1 ASummKLAInpKLSummInpJS

SummaryandInputofondistributimeanSummInp

A ,2

A range of distributional similarity and other

features◦ KL divergence

◦ JS divergence

◦ Cosine similarity

◦ Topic signature based features

◦ Summary likelihood under a frequency based model

11

12

Pyramid Resp.

JS divergence -0.74 -0.71

Pyramid Resp.

-0.72 -0.61

Query Task Update Task

Regression 0.77 0.67 0.71 0.54

Good content can be predicted from the input

Information-theoretic features provide good

estimates

Best performance on ‘08: JS = 0.89 with pyramid scores

Average system performance on an input can be

predicted with good accuracies

13

Systems ignore properties of individual inputs◦ Very low average performance on certain inputs

14

Input difficulty can be measured by a number of

features [ACL ‗08, EMNLP ‗09]

Can predict when average system performance

will be below the mean value

15

Difficult input◦ Most systems perform poorly

◦ Low average system score

2 classes – easy, difficult◦ Above/below mean average system score

◦ Equal number of inputs in both classes

◦ Large vocabulary size

◦ Fewer descriptive words – hard to identify through

frequency and repetition

◦ Low redundancy between input documents

◦ No clear topic

16

6 significant features

Good accuracies in identifying difficult inputs

10% above baseline

17

All inputs Extremes

Query task 61.36 60.00

Update task 59.09 75.00

Properties of input predictive of average system

performance

Specialized content selection necessary to smooth

out variations

* Extremes –

10 each most

easy and difficult

Trained on DUC

2002-2004

Pseudo-models for summary evaluation

18

Another likely setup on non-standard test sets

Robust system-level rankings on large test sets

RSU4-recall Pyramid Resp.

1 model 0.92 0.80

4 models 0.92 0.79

Query Task

Pyramid Resp.

0.80 0.69

0.85 0.69

Update Task

Choose one model per input ◦ Alphabetical order of model name

Considerably fewer inputs with significant

correlations

RSU4-recall Pyramid Resp.

1 model 84.09 79.54

4 models 95.45 81.82

Query Task

Pyramid Resp.

86.36 75.00

100 86.36

Update Task

Related work in Machine translation

One human reference translation

Off-the-shelf systems as pseudo-references

Features to compare other translations with

pseudo-references

Regression based scoring

Improved correlations compared to using a single

human reference

22

Pseudo-model ~ systems predicted to be best

using available model summary

Compute ranks based on the human model

Treat top systems as ―pseudo-models‖

Global◦ System level ranking using RSU4

◦ Select top 3 systems as pseudo-models

Local◦ Use top 3 systems for each input as pseudo-models

Final rankings◦ JS divergence with 1 model + 3 pseudo-models

Improvements for pyramid

Not much gains for responsiveness

On ‘08 data, local selection was better

Global sel. 93.18 79.55

Local sel. 93.18 75.00

93.18 77.27

86.36 79.55

Query Task Update Task

Pyramid Resp.

1 human 84.09 79.54

Pyramid Resp.

86.36 75.00

+ 3

pseudo

Collection of system summaries is useful for

evaluation

25

Similar to the pyramid method ◦ Common content across multiple human summaries

more important

Different systems ~ different content selection

methods◦ Agreement among systems ~ very important content

Collection of system summaries as a model ◦ Indicative of what is important?

Divergence from vocabulary distribution of system

summaries

Collective

vocabulary of all

system summaries

Vocabulary

distribution of

individual system

summary

Low divergence ~ higher scores

System-level Pyramid Resp.

System summaries -0.93 -0.81

Pyramid Resp.

-0.89 -0.79

RSU4 – 4 models 0.92 0.79 0.85 0.69

Query Task Update Task

Percentage of inputs with significant correlations◦ 77 to 90%

Collective knowledge of systems is useful◦ Possibility of system combination for summarization

Based upon system summaries◦ Pseudo-models: help only for pyramid correlations

◦ Collection of system summaries: very indicative of good

content

29

4 methods to predict summary quality that use

very little or no human input

Based upon the input◦ Input-summary similarity: highly predictive

◦ Input difficulty features: predictive of average system

performance

Automatically Evaluating Content Selection in

Summarization without Human Models◦ Annie Louis & Ani Nenkova, EMNLP 2009

Performance Confidence Estimation for Automatic

Summarization◦ Annie Louis & Ani Nenkova, ACL 2009

Summary Evaluation without Human Models◦ Annie Louis & Ani Nenkova, TAC 2008

Can you summarize this? Identifying correlates of input

difficulty for generic multi-document summarization◦ Ani Nenkova & Annie Louis, ACL-HLT 2008

30